The neural circuitry of fear conditioning : a theoretical account

(1)

HAL Id: tel-01780849

https://tel.archives-ouvertes.fr/tel-01780849

Submitted on 28 Apr 2018

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of

sci-entific research documents, whether they are

pub-lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diffusion de documents

scientifiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

The neural circuitry of fear conditioning : a theoretical

account

Martin Angelhuber

To cite this version:

Martin Angelhuber. The neural circuitry of fear conditioning : a theoretical account. Neurons and

Cognition [q-bio.NC]. Université de Strasbourg; Albert-Ludwigs-Universität (Freiburg im Breisgau,

Allemagne); Universität Basel, 2016. English. �NNT : 2016STRAJ082�. �tel-01780849�

(2)

UNIVERSITÉ DE STRASBOURG

ÉCOLE DOCTORALE DES SCIENCES DE LA VIE ET DE LA SANTÉ

Institut des neurosciences cellulaires et intégratives (INCI)

THÈSE

présentée par :

Martin ANGELHUBER

soutenue le 27

octobre 2016

pour obtenir le grade de :

Docteur de l’université de Strasbourg

Discipline/ Spécialité

: Neurosciences computationnelles

The Neural Circuitry of Fear

Conditioning

A Theoretical Account

Thèse en cotutelle avec les universités de Fribourg (Allemagne) et Bâle (Suisse)

Programme NeuroTime

THÈSE dirigée par :

M. Ad Aertsen

Albert-Ludwigs-Universität Freiburg

M. Andreas Lüthi

Universität Basel

M. Pierre Veinante

Université de Strasbourg

RAPPORTEURS :

M. Markus Diesmann

Forschungszentrum Jülich

M. Cyril Herry

Université de Bordeaux

AUTRES MEMBRES DU JURY :

M. Arvind Kumar

KTH Stockholm

M. Stefan Rotter

Albert-Ludwigs-Universität Freiburg

(3)

The Neural Circuitry of

Fear Conditioning

A Theoretical Account

Th`

ese en cotutelle

pr´

esent´

ee par

Martin Angelhuber

pour obtenir le grade de Docteur de

l’universit´

e de Fribourg-en-Brisgau,

l’universit´

e de Bˆ

ale et

l’universit´

e de Strasbourg

Discipline/Sp´

ecialit´

e: Sciences de la Vie/Neurosciences

´

Ecole Doctorale Sciences de la Vie et de la Sant´

e (ED 414)

Institut des Neurosciences Cellulaires et Int´

egratives (CNRS UPR

(4)

ii

This thesis summarizes research performed at the Bernstein Center Freiburg, the Friedrich Miescher Institute Basel and the Université de Strasbourg within the Erasmus-Mundus-Joint-Doctorate-Programme “NeuroTime”. Accordingly, it has also been submitted to the Faculty of Biology of the Albert-Ludwigs-Universität Freiburg and the Faculty of Philosophy and Natural Sciences of the Universität Basel.

(5)

Abstract

In the last decades, fear conditioning has been established as one of the most successful paradigms for studying the neural substrates of emotional learning. Experimental research has revealed a complex circuitry of brain regions—most prominently the amygdala—underlying the acquisition, extinction and general-ization of conditioned fear. As the wealth of experimental data grows, theoretical models that help interpret results and generate new hyoptheses play an increas-ingly important role. In this thesis, two computational models of the neural substrates of fear conditioning are presented.

The first model is a biologically realistic spiking neural network model of the central amygdala, the main output structure of the amygdala. Based on a recent experimental study that demonstrated the importance of tonic extrasynaptic inhibition for fear generalization, the effects of changes in neuronal membrane conductance on input processing are analyzed in the model. Consistent with experimental results, it is shown that subpopulation-specific changes in tonic inhibitory conductance increase the responsiveness of the network to phasic inputs, presumably causing the increase in fear generalization. On the basis of this result, the model is analyzed from a functional perspective. It is argued that tonic inhibition in the central amygdala acts as a controller by which network sensitivity is flexibly adjusted to relevant features of the environment, such as predictability of threat, and concrete predictions that follow from this proposition as well as possible adjustment mechanisms are discussed.

In addition, a systems level model is presented that is based on a recent high-level approach to conditioning and proposes a specific physiological imple-mentation in the basolateral amygdala, prefrontal cortex and the intercalated cell clusters of the amygdala. It is a central hypothesis of the model that the interaction between fear and extinction neurons in the basal amygdala, which has been described experimentally, is a neural substrate of the switching between socalled latent states, which allow the animal to organize its experience and infer structure in the environment. Important behavioral phenomena are reproduced in the model and the effect of de-activation of model structures is shown to be in good agreement with results from lesion studies. Finally, predictions and questions that follow from the main hypothesis are considered.

Taken together, the two models provide a coherent theoretical account of the neural basis of acquisition and extinction of conditioned fear, as well as the control of fear generalization. Importantly, this account combines different levels of analysis. By virtue of this combination, the scope of predictions that can be derived is expanded and the models become more amenable to experimental testing.

(6)

iv ABSTRACT

List of Publications

The following articles include work presented in this thesis and are prepared for publication.

• Tonic Inhibition Controls Fear Generalization in a Network Model of the Central Amygdala.

Martin Angelhuber, Paolo Botta, Andreas L¨uthi, Ad Aertsen, Arvind Kumar

Chapter 5 of this thesis.

• A Computational Model of State-Switching in the Basal Amygdala during Fear Learning.

Martin Angelhuber, Andreas L¨uthi, Ad Aertsen, Arvind Kumar Chapter 6 of this thesis.

• A Fokker-Planck approximation for conductance-based IAF neurons and its application to the analysis of inhibitory networks.

Martin Angelhuber, Ad Aertsen, Arvind Kumar Sections 4.2 and 4.3 of this thesis.

• Spatial architecture generates bumps of activity and input-dependent dynamics in purely inhibitory networks

Sebastian Spreizer, Martin Angelhuber, Jyotika Bahuguna, Ad Aertsen, Arvind Kumar

(7)

Zusammenfassung

Angstkonditionierung hat sich in den letzten Jahrzehnten als eine der erfolgreich-sten Methoden zur Untersuchung der neuronalen Substrate von Emotionslernen etabliert. Experimentelle Forschung hat ein komplexes Netzwerk verschiedener Hirnstrukturen, das dem Erwerb, der Extinktion und der Generalisierung kon-ditionierter Angst zugrunde liegt und in dem die Amygdala eine Schl¨usselrolle einnimmt, aufgedeckt. Da die Menge an experimentellen Daten immer st¨arker zunimmt, kommt theoretischen Modellen, die der Einordnung experimenteller Ergebnisse und dem Aufstellen neuer Hypothesen dienen, eine immer gewichtigere Rolle zu. In dieser Dissertation werden zwei theoretische Modelle zu den neu-ronalen Substraten von Angstlernen vorgestellt.

Bei dem ersten Modell handelt es sich um ein biologisch realistisches Netz-werkmodell mit spikenden Neuronen, das der zentralen Amygdala nachempfunden ist. Auf Grundlage einer experimentellen Studie, die einen Zusammenhand zwischen extrasynaptischer Inhibition und Angstgeneralisierung demonstriert hat, werden die Folgen von ¨Anderung der neuronalen Membranleitf¨ahigkeit auf die Informationsverarbeitung im Gesamtnetzwerk analysiert. Dabei wird gezeigt, dass—im Einklang mit experimentellen Ergebnissen—populationsspezifische

¨

Anderungen die Ansprechempfindlichkeit des Netzwerks maßgeblich erhöhen. Ausgehend von diesem Ergebnis wird das Modell einer funktionalen Analyse unterzogen. Es wird vorgeschlagen, dass extrasynaptische Inhibition in der zentralen Amygdala als Regler fungiert, mit Hilfe dessen Netzwerksensitivität flexibel den Begebenheiten der Umwelt, wie z.B. Vorhersagbarkeit von Gefahr, angepasst werden kann, und konkrete Vorhersagen, die aus dieser Hypothese folgen, sowie mögliche Mechanismen, werden erörtert.

Des weiteren wird ein Modell auf Systemebene präsentiert, das auf einem kürzlich vorgeschlagenen Konditionierungsmodell aus den Kognitionswissen-schaften aufbaut und eine physiologische Implementierung in der basolateralen Amygdala und dem präfrontalen Kortex untersucht. Die Grundannahme des Modells ist, dass die Wechselwirkung zwischen Angst- und Extinktionsneuronen in der basalen Amygdala, die experimentell beschrieben wurde, ein neuronales Substrat des Umschaltens zwischen latenten Zuständen ist, die es dem Tier ermöglichen seine Wahrnehmungen zu organisieren und Strukturen in der Umwelt zu erkennen. Das Modell reproduziert wichtige Verhaltensphänomene und die Folgen von Manipulationen im Modell sind in gutem Einklang mit den Folgen von Läsionen der entsprechenden Hirnregionen. Darüberhinaus werden die Vorhersagen und offenen Fragen, die sich aus der Grundhypothese ergeben, diskutiert.

Zusammen bilden die beiden Modelle eine koh¨arente Beschreibung von Erwerb und Extinktion konditionierter Angst und der Regelung von Angstgeneralisierung.

(8)

vi ZUSAMMENFASSUNG

Diese Beschreibung kombiniert verschiedene Analysebenen. Durch diese Kombi-nation erweitert sich die Möglichkeit Vorhersagen abzuleiten beträchtlich und die Modelle werden experimenteller Untersuchung zugänglich.

(9)

R´

esum´

e

Au cours des dernières décennies, le conditionnement à la peur a été établi comme un des paradigmes les plus réussis pour comprendre les substrats neuronaux de l’apprentissage et de l’émotion. La recherche expérimentale a révélé les structures du cerveau, plus importante l’amygdale, qui sous-tendent l’acquisition, l’extinction et la généralisation de la peur conditionnée. Comme la richesse des données expérimentales ne cesse de croˆıtre, des modèles informatiques peuvent aider à interpréter les résultats et contribuer à notre compréhension du circuit neural du conditionnement à la peur. Dans cette thèse, je présente deux modèles informatiques à cet effet.

Le premier modèle est un modèle biologiquement réaliste de l’amygdale centrale simulant un réseau de neurones en activité. Sur la base des études récentes reliant l’inhibition tonique et la généralisation de la peur, le modèle est utilisé pour enquêter sur l’effet des changements de l’inhibition tonique sur le traitement des informations re¸cues. L’analyse confirme que la diminu-tion de l’inhibidiminu-tion tonique d’une populadiminu-tion augmente la réactivité du réseau aux informations phasiques re¸cues. Ce résultat est cohérent avec les résultats expérimentaux et corrobore le lien entre l’inhibition tonique et la généralisation de la peur précédemment décrite. Ensuite, le modèle est analysé d’une perspec-tive fonctionelle. On propose que l’inhibition tonique agit comme un régulateur pour ajuster la réactivité à un certain nombre de facteurs, principalement la prévisibilité du stimulus inconditionnel. Des prédictions qui découlent de cette proposition ainsi que des mécanismes d’ajustement possibles sont discutés.

En outre, je présenterai un modèle systématique, centré sur l’amygdale baso-latérale contenant le cortex préfrontal et les cellules intercalées de l’amygdale. Ce modèle est basé sur un type de modèle de conditionnement récemment introduit dans les sciences cognitives utilisant des variables latentes pour re-connaˆıtre la structure de l’environnement et prédire le stimulus inconditionnel. C’est une hypothèse centrale du modèle que l’interaction entre les neurones de la peur et les neurones d’extinction dans l’amygdale basale, qui ont été décrits expérimentalement, code pour l’interface entre les variables latentes. Sur la base de cette hypothèse, il est démontré que le modèle couvre une large gamme d’effets, commen¸cant par des effets purement comportementaux jusqu’aux résultats d’études lésionnelles. De plus, l’analyse du modèle produit un certain nombre de prédictions vérifiables qui seront discutées en détail.

Pris ensemble, les deux modèles offrent une perspective théorique cohérente de la base neurale de l’acquisition et de l’extinction de la peur conditionnée, ainsi que le contrôle de la généralisation de la peur. Cette approche combine des niveaux d’analyse différents. De cette fa¸con, plus de prédictions peuvent être dérivées et les modèles se prêtent mieux à des tests expérimentaux .

(10)

Abbreviations and Symbols

Conditioning Terminology

FC classical fear conditioning CS conditioned stimulus US unconditioned stimulus CR conditioned response UR unconditioned response RPE reward-prediction error TD temporal-difference

RLSC reinforcement learning and state classification PREE partial reinforcement extinction effect

Anatomy

LA lateral amygdala BA basal amygdala

BLA basolateral complex of the amygdala CEA central amygdala

CEl lateral part of the central amygdala

CElon CEl subpopulation innervated by CS after conditioning (see 2.3.2) CEloff CEl subpopulation inhibited by CS after conditioning (see 2.3.2) CEm medial nucleus of the central amygdala

ITC intercalated cell cluster mITC medial intercalated cell cluster mPFC medial prefrontal cortex IL infralimbic cortex PL prelimbic cortex HPC hippocampus

BNST bed nucleus of the stria terminalis PAG periaqueductal grey

Neurochemicals

GABA γ-Aminobutyric acid NMDA N-Methyl-D-aspartate

AMPA α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid SOM somatostatin

PV parvalbumin PKC protein kinase C

CRF corticotropin-releasing factor

(13)

xi

Mathematical Notation and Symbols

x, X italic scalar x boldface vector X upright capital letter matrix

E expectation value V variance

N (µ, C) Gaussian distribution B(a, b) Beta distribution G(n, θ) Gamma distribution F T [x] orx_e Fourier transform of x

(14)

(15)

Chapter 1 Introduction

Throughout their lives, all animals, including humans, navigate a delicate trade-off: On the one hand, predicting potential threat and reacting appropriately is obviously crucial for survival. On the other hand, excessive fear and anxiety are clearly detrimental to other behaviors critical for evolutionary fitness, and, in the case of humans, severely impair quality of life. To keep this balance in an ever-changing environment, animals rely on learning mechanisms that allow them to adapt to novel threats.

In recent decades, neurobiological research has begun to reveal the neural substrates of such behavioral adaptations in rodents. A quickly expanding catalog of experimental studies maps the neural circuitry of fear learning in ever greater detail and an intricate arrangement of a number of brain structures emerges, with the amygdala taking center stage. As the complexity of this circuitry becomes increasingly apparent, the need for theoretical interpretation only becomes more urgent.

1.1 Aim of the Thesis

With this work, I endeavour to contribute to this ongoing research effort by proposing a theoretical account of the neural circuitry of fear learning. In particular, two computational models are presented in this thesis.

The first model is a biologically realistic spiking neural network model of the central amygdala, which is closely based on experimental data and examines the role of tonic inhibition in controlling fear generalization from both a mechanistical and functional perspective. It corroborates recent experimental findings on the relation of tonic inhibition and fear generalization and expands on the role of the central amygdala in fear expression, or, more generally, action selection.

The second model, is based on a recent high-level approach to conditioning

(16)

2 CHAPTER 1. INTRODUCTION

using latent variables, itself grounded in the theory of Bayesian inference. With this as a starting point, a physiologically constrained implementation is developed and analyzed. The resulting model yields an explanatory framework for a wide number of experimental results and makes hypotheses on the roles of many structures which have been found to be implicated in fear. Both of these models allow for a number of testable predictions that are discussed in detail.

Furthermore, as a tool to help implement and interprete spiking neural network simulations, an analytical approximation to the mean firing rates of the conductance-based integrate-and-fire neuron model has been derived, using the Fokker-Planck formalism for diffusion problems. This approximation is used for analyzing the dynamics of inhibitory networks.

In this thesis, I try to bring together different approaches to studying fear conditioning theoretically. It is my hope, that it contributes towards bridging the gap between high-level models of conditioning, solely based on behavior, and biologically realistic neural network models, based on neurophysiological data. As a consequence, many of the predictions and hypotheses derived from this work argue for increasingly combining setups used in behavioral studies with more recently available neurophysiological measurements and manipulations.

1.2 Classical Fear Conditioning

Classical conditioning was first described by Ivan Pavlov (Pavlov, 1927) and has since become one of the most important experimental paradigms to study learning in animals. In classical conditioning, an initially neutral stimulus is paired repeatedly with an appetitive or aversive stimulus. As a result, the neutral stimulus comes to evoke a response as well.

1.2.1 Experimental Procedure

Before the main phase of the experiment, the animal is allowed time to get used to the location in which the conditioning will occur, a phase referred to as habituation. Then, in the actual training phase, an initially neutral stimulus, usually a tone or light, is paired repeatedly with the unconditioned stimulus (US). The US is a stimulus with clear motivational valence, i.e., clearly appetitive or aversive. As a consequence of this pairing, the animal acquires responses to the initially neutral stimulus. These responses are termed conditioned responses (CR), since their appearance is conditional on the previous acquisition, and, correspondingly, the stimulus evoking them is called conditioned stimulus (CS). In the case of fear conditioning, the US is most often a painful electric shock, either to the paws or eyelids; and the conditioned response is typically freezing, a

(17)

1.2. CLASSICAL FEAR CONDITIONING 3

brief period of immobility but may also comprise changes in heart rate, analgesia, and release of stress hormones (LeDoux, 2000).

Timing of CS and US

What is meant exactly with pairing in the previous paragraph merits further clarification. If the CS and US overlap entirely, i.e., they start and end at the same time, we speak of simultaneous conditioning. More commonly, however, the US presentation begins after the CS onset. Depending on the relative timing of CS-ending and US-beginning, two cases can be distinguished. In delay conditioning, the US begins before or immediately when the CS ends. In trace conditioning, on the other hand, the US onset is after the ending of the CS, and the temporal gap between the two stimuli is referred to as trace interval (Bouton,

2007). The different temporal arrangements can lead to different results. The longer the gap between CS and US, the harder it is to learn the association and with more than a few seconds of trace interval, no learning is achieved at all (Smith, 1969). Another important example for the criticality of timing is the difference between second-order conditioning and conditioned inhibition, which will be explained later. It is outside the scope of this work to elaborate on these effects in detail; all the results should be understood as pertaining to delay conditioning with the US directly following the CS. This is the procedure most commonly used in the experiments the work is based on.

Discriminative Conditioning

For many purposes, it is useful to introduce an additional control stimulus, e.g., a tone of a different frequency, which is also presented during training, but not paired with the US. To indicate it was not paired, the superscript “-” will be used, as opposed to the CS+, the conditioned stimulus that was actually paired. Whenever more than one CS+ or CS− is used, we use subscripts to denote stimulus idendity. For instance, stimuli CS+₁ and CS+₂ would be two different stimuli that were both paired with the US.

After the training phase, the persistence of acquired responses is verified in the next phase. This phase is often performed in a different context, e.g., a markedly different cage, to confirm the response is CS- and not context-specific. In the testing phase, the CS is typically not paired with the US. If the study involves extinction learning, the CS is presented repeatedly without the US in this phase, leading to a slow decline in conditioned responding. In this case, a separate testing phase is executed after extinction learning, often back in the original conditioning context.

(18)

1.2.2 Extinction Learning and Fear Generalization

A pertinent observation about extinction learning is the instability of the extinc-tion memory, meaning that condiextinc-tioned responses reappear occasionally. This can be triggered by a number of manipulations and the effect is termed accord-ingly: Renewal describes the renewed emergence of conditioned responses when switching to a novel or the training context (Bouton,2004). This effect points towards the high context-specificity of extinction memory. Another way to renew conditioned responding is to present the US alone, this is termed reinstatement (Rescorla, 1975). In addition to these two, conditioned responding could also reappear spontaneously, in which case it is termed spontaneous recovery (see figure 1.1).

The multitude of extinction effects already points towards an important advantage of classical conditioning: simple as the paradigm might be, there is a wealth of experimental variations that are possible within its boundaries and lead to effects that can shed light on a wide range of learning mechanisms. Many of the variations used in neurobiological settings focus on the study of fear gen-eralization and fear extinction, two aspects of learning that are of high relevance to pathological behavior. More precisely, the exact readout for quantifying fear extinction is the exhibition of the conditioned response, i.e., freezing rates, in the testing phase. Fear generalization is typically quantified by the ratio of CS− to CS+response rates. A high ratio indicates that the animal does not discriminate between CS− and CS+. More generally, in studying stimulus generalization in conditioning, it is found that conditioned responding to the CS− depends on similarity. When plotted along a sensory continuum, e.g., tone frequency in the case of auditory conditioning, conditioned responding is maximal at CS+ and decreases as similarity decreases, yielding a bell-shaped generalization curve (Pavlov,1927). Remarkably, these generalization curves stretch over perceptual boundaries, e.g., between colors (Guttman,1956). This indicates that stimulus generalization is more than a mere failure at sensory discrimination; it includes an active cognitive component (Shepard, 1987; Dunsmoor,2015).

1.2.3 Variations of the Paradigm and Notable Effects

Complementing the standard paradigm is a number of experimental variations that allow for investigation of a wide range of effects. These have so far mostly been employed in animal psychology studies—some in appetitive conditioning— and contributed greatly to the development of behavioral models of conditioning. While they have so far mostly been restricted to setups without recordings of neural activity, it is to be expected that, as recording techniques improve, they can be used in conjunction with recording of neural activity in the near future

(19)

Figure 1.1: Classical fear conditioning. During training, the animal ac-quires a conditioned response (freezing) due to the repeated pairing of the CS (tone) and US (footshock). Afterwards, during extinction learning, the freezing response diminishes as the CS is presented without the US. Here, a discrimina-tive paradigm in which a second tone (CS−) is presented during training but not paired with the US is depicted. On the right side, different modes of CR re-occurence are sketched: renewal, which is caused by change of context; rein-statement, in which the CR returns after a single unpaired US; and spontaneous recovery, where the CR re-occurs after some time.

to add to our understanding of the neural circuitry.

Second-Order Conditioning and Sensory Preconditioning

There are two noteworthy variations demonstrating that a CS can elicit a response even though it has never been paired with the US itself. Firstly, in second-order conditioning, a CS (CSA) is directly followed by the US in the first phase of

the experiment. In a second phase, this CSAis presented right after a different

CS (CSB). Remarkably, CSB also acquires a response (e.g. Gewirtz, 2000),

demonstrating that a conditioned stimulus can itself act as a reinforcement signal after learning.

This can be taken even further in sensory preconditioning (Bouton,2007): CSA and CSB are paired in the first phase of the experiment. In the second

phase, stimulus CSA is paired with a US. Consequently, CSB also elicits a

conditioned response in the testing phase. Again, CSB has never been paired

with US. Notably, though, in sensory preconditioning—unlike second-order conditioning—it also never co-occured with the conditioned response before testing. This strongly implies that associations are formed between stimuli rather than stimulus and response and that already motivationally irrelevant stimuli, such as the two CSs before learning, do form these associations.

(20)

Latent and Conditioned Inhibition

More evidence for learning processes in the absence of an US comes from a phenomenon termed latent inhibition (Lubow, 1965). Here, stimulus CSA is

presented repeatedly without the US in the first phase. When it is then paired with the US in the second phase, the acquisition of a conditioned response is significantly delayed. This indicates stimulus-specific learning in the first phase of the experiment without US presentations.

Similarly, a stimulus can be trained to inhibit conditioned responding to other stimuli (Rescorla, 1969). If a previously conditioned stimulus CSA is

paired with stimulus CSB in the absence of the US, CSB reduces conditioned

responding when presented together with other previously conditioned stimuli, an effect referred to as conditioned inhibition. Note the strong similarity of this paradigm with second-order conditioning. This example highlights how critical exact timing between the stimuli is: A subtle difference in relative timing can lead to diametrically opposite effects. Nonetheless, usually both learning processes—second-order conditioning and acquisition of conditioned inhibition— develop simultaneously, with a tendency for second-order conditioning to be acquired a bit faster. This leads to an overall non-monotonic learning curve and greatly complicates the interpretation of results (seeGewirtz,2000;Yin,1994).

Cue Competition Effects

The previous examples already included schedules with more than one CS and demonstrated that these stimuli mutually interact in forming US associations. Cue competition effects are a specific class of phenomena with multiple CSs in

which the CSs compete for association with the US. The most prominent of these is Kamin blocking (Kamin,1969). In Kamin blocking, a previously conditioned CS (CSA) is paired with CSB and the US in the second phase of the experiment.

As a consequence of the pairing with CSA, CSB acquires no, or a much weaker,

response than a suitable control. Importantly, Kamin blocking was a key insight and motivation behind the formulation of the Rescorla-Wagner model described later.

Other cue competition effects include overshadowing, in which two CSs are paired with the US, and depending on factors like salience, one of them acquires a much stronger response than the other, and relative validity (Wagner,

1968). Here, three distinct stimuli, CSA, CSB and CSX, are involved and during

conditioning both CSA and CSB are always paired with CSX, i.e., compounds

CSAX and CSBX are used. In one group of subjects, CSAX is always presented

together with the US, while CSBX is always presented without the US. In the

(21)

Interestingly, even though CSX is paired equally often with the US in both

groups, it elicits a significantly stronger response in the latter group. This finding highlights the importance of US prediction (Rescorla,1988): In the first group, CSA is a much better predictor of the US than CSX and accordingly

acquires a strong response at the expense of CSX. In the second group, however,

all three stimuli are equally predictive of the US, since all of them were paired with the US half of the time.

Occasion Setting and Configural Conditioning

So far, only the linear interaction of stimuli was considered, i.e., each CS was either a conditioned excitor (increasing the response probability) or inhibitor (decreasing it) and the response to the presentation of both of them together could be considered the sum of their individual effects. There are, however, many cases in which the interaction between stimuli is nonlinear. One specific case is called occasion setting (Holland,1989;Bouton,2007), in which a third stimulus merely modulates the association between a given CS and the US. Consider the example of feature-positive discrimination: stimulus B always precedes CSA

whenever CSA is paired with the US, but not when it is presented alone. The

animal can learn that CSA is predicting the US only when B was also presented.

Importantly, B does not act as an excitor; when presented with a third stimulus it has no effect, i.e., it very specifically modulates the association between CSAand

the US. Conversely, in feature-negative discrimination, the occasion setter signals the absence of the US. These findings point towards hierarchical organization of learning processes, where learning the role of stimulus B is specific to the CSA-US association.

Partial Reinforcement

Finally, another often used variation is conditioning with partial reinforcement, i.e., not every presentation of the CS is accompanied by the US. There is a variety of schedules, some deterministic (e.g., only every other CS is paired with the US), and some random (e.g., CS and US are paired with 50% probability). Usually, either the length of the acquisition phase is adjusted or unpaired US presentations are added, such that the overall reinforcement during training is the same as in the fully conditioned control group (Haselgrove,2004). Irrespective of the exact schedule, a very salient and robust finding is the partial reinforcement extinction effect, the observation that extinction learning after partial conditioning is delayed as compared to the fully conditioned control animals (Haselgrove,2004;Gallistel,

2000). Importantly, this contradicts the traditional associative account that conditioned responding reflects the strength of the association between the CS

(22)

and US.

Taken together, this wide range of effects illustrates the wealth and informa-tive value of this seemingly simple paradigm. Many of the described phenomena demonstrate that the mere temporal co-occurence of CS and US is neither sufficient nor necessary for the acquisition of a CR. Evidence has accumulated that a computational framework that conceptualizes conditioning as the attempt at predicting US occurence based on previous experience provides a better fit to empirical data compared to mere associative learning between coinciding stimuli (Rescorla,1988). Accordingly, throughout the last decades, theoretical models and interpretations of conditioning have been developed based on these observations. These will be discussed in chapter three.

1.3 Fear and Anxiety

The prior discussion focussed on conditioning per se, and was not specific to fear or anxiety. Here, these terms are introduced in more detail. Importantly, while the two terms are often used almost interchangeably in colloquial discourse, a clear distinction is made in technical language. Fear refers to an acute defensive reaction against a specific perceived threat, whereas anxiety is a sustained and general mood of vigilance and unease linked to the vague anticipation of future negative events (see e.g.Davis,1992). For animal research, the notions of fear and anxiety are linked to observable behaviors in standard paradigms.

1.3.1 Fear in Animals

The gold standard for studying fear is the previously described paradigm of classical fear conditioning. As it is not possible to make meaningful claims about the emotional experiences of animals, fear is simply a theoretical construst underlying the observed responses (Davis,1992). In the school of operational behaviorism, it can be conceived as an intervening variable, a variable that might not be directly observable variable, and that combines a possibly diverse list of stimuli and responses into a coherent explanation of behavior (see Figure 1.2 andBouton(2007);LeDoux(2014)). Note that in this scheme, the intervening variable is linked to both stimuli and responses, and these links make the system in principle falsifiable. For all practical purposes, however, the observable responses themselves, like freezing and startle, are more commonly taken to define fear in a specific experimental setting. Nevertheless, when viewed as an intervening variable, fear could be given a definition that goes beyond freezing and that still lives up to the standards of scientific rigor. This subtle difference underlies some theoretical considerations that are discussed later. For now, it

(23)

1.3. FEAR AND ANXIETY 9

Figure 1.2: Fear and anxiety. a) Fear as an intervening variable creates a conceptual link between a number of observables. b) Important behavioral assays for testing anxiety: The open field test (left) in which the relative time spent in the center (red square) is used as an inverse measure of anxiety, and the elevated plus maze (right) in which the relative time spent in the open arms is used to quantify anxiety.

suffices that the notion of fear is inextricably linked to observable behavior.

1.3.2 Animal Models of Anxiety

Similarly, the notion of anxiety also relies on observable behaviors in experimental tests. The two tests most commonly used are the elevated plus maze (Pellow,

1985) and the open field test (Hall,1932;Denenberg, 1969;Carola,2002). Both tests exploit the balance between two opposing natural urges rodents display: exploration and defensive avoidance (Blanchard,2008;Tovote,2015). On the one hand, rodents have a natural tendency to explore their environment, but on the other hand, they tend to avoid open spaces and possible exposure to predators. In a big open field, as well as in a plus maze in which only two arms are sheltered (see figure 1.2), these two tendencies conflict with each other. As a consequence, behavior is very sensitive to the sustained mood of the animal. A pertinent observation is that animals that have undergone fear conditioning or other putatively traumatic experiences are more likely to avoid open spaces. Hence, they tend to stay close to the walls in the open field test, or within the sheltered arms in the elevated plus maze. The relative time spent in the open spaces can be used as an inverse quantifier of anxiety: The more time spent in the open, the less anxious the animal. Notably, this quantifier has also been shown to be sensitive to the application of anxiolytic drugs (Pellow, 1986;

(24)

1.3.3 Relation between Fear and Anxiety

From early on, theoretical accounts of anxiety have implicated conditioning in the emergence of anxiety disorders (Watson, 2002; Pavlov, 1927). Some disorders, like post-traumatic stress disorder, are often conceptualized within the conditioning framework as deficits in extinction learning and overgeneralization of fear. Accordingly, anxiety disorders are much more prevalent among combat and trauma survivors (Dohrenwend,1981;Lissek,2005). On the other hand, one of the main criticisms of this conditioning model of anxiety in humans is that very often there is no relevant history indicating conditioning-like mechanisms in people with phobias (Rachman,1990). Still, as more complex conditioning phenomena were discovered, it was argued that many observations on the emergence of anxiety disorders, which seemed to be at odds with the idea of a direct link between fear learning and anxiety, can be explained in terms of these phenomena (Mineka,2006). For instance, latent inhibition can account for between-individual differences in reactions towards traumatic events, depending on their previous experience with the stressor; second-order conditioning or vicarious conditioning1 can explain how phobias can form without explicit pairing with an aversive event. Finally, the conditioning model of anxiety is also validated by the sucess of exposure therapy for the treatment of pathological anxiety (Barlow,2002).

This is, of course, not to understate the importance of other individual factors, like genetic predisposition. Still, there is broad consensus that the study of conditioning phenomena can inform our understanding of the emergence of anxiety and anxiety disorders. Here, some theoretical considerations and empirical evidence on the link between fear and anxiety are presented.

Deficits in Extinction Learning

The conditioning model of anxiety proposes that pathological anxiety rests on a failure to extinguish previously acquired conditioned responses (Eysenck,

1979;VanElzakker,2014). Overall, the empirical evidence supports that anxiety disorders are associated with heightened conditioned responding during extinction learning (Lissek,2005;Blechert, 2007;Peri,1999) and also during extinction recall (Milad, 2008,2009). Importantly, this relationship between anxiety and resistance to extinction learning could be reproduced in rodents by breeding selectively high- and low-anxiety rats (Muigg,2008). In addition, concomittant measurements of neural activity confirmed the involvement of the fear extinction circuitry for this process (ibid.).

1_{Vicarious conditioning names to the phenomenon that individuals can acquire fear responses}

to a CS by observing other conspecifics’ fearful reaction to that CS. This can be shown to occur in, e.g., rhesus monkeys (Cook,1989).

(25)

1.3. FEAR AND ANXIETY 11

Fear Generalization

Generalizing the above ideas on extinction, recent theories link anxiety to a failure to inhibit fear responses during safety learning (Davis,2001;Jovanovic,

2012). In line with this, increased CR rates (as compared to healthy controls) on CS− presentation, i.e., higher fear generalization, have been reported in anxiety patients in a number of studies (Grillon,1999;Peri,1999;Glover,2011;

Dunsmoor, 2015). In addition, studies in rodents revealed a consistent relation between inter-individual differences in fear generalization scores and anxiety (Duvarci,2009;Botta,2015): Animals that displayed high fear generalization

also tended to score high on anxiety tests.

US-Predictability

Finally, an important finding on the nature of sustained fear and anxiety is that unpredictable aversive events are much more likely to lead to sustained fear (Davis,2010;Walker, 2009). When comparing two groups of subjects—one which underwent classical conditioning with CS-US pairing and another in which both stimuli were presented equally often but not paired with each other—it is found that the latter displays much higher sustained fear, while the first only exhibits phasic and CS-specific fear responses (Davis,2010). This is consistent with contemporary interpretations of conditioning as US prediction: In case the CS is a clear predictor, no strong associations are formed with contextual cues; but in case there are no phasic predictors, contextual cues form US presentations, resulting in sustained and rather undirected states of fear. More generally, the idea that uncertainty about future threats results in anxiety and that maladaptive responses to uncertainty underly many disorders is central to a recently proposed anxiety model (Grupe,2013).

In summary, these results demonstrate a link between fear learning and the emergence of anxiety. More particularly, two specific facets of this link should be highlighted: Firstly, the emergence of anxiety depends crucially on predictability. Anxiety is more likely to develop whenever the environment does not allow for the prediction of aversive events, thus undercutting the ability to avoid them or extenuate their effect. Secondly, sustained fear, or anxiety, is related to the expression of phasic fear. Hypersensitivity to phasic cues, as in the above examples of extinction learning and fear generalization, is usually considered a hallmark of anxiety (Blanchard,2008). These two aspects provide the foundation for relating results of the conditioning models to anxiety in later chapters.

(26)

1.4 Fear as a General Model for Learning

Apart from its high clinical relevance for the study of pathological anxiety, it deserves emphasis that fear conditioning is a highly attractive model for studying learning in general. It provides noteworthy practical advantages, stemming from the very nature of fear learning and common to all variations of the paradigm: Firstly, there are clear, quantifiable behavioural readouts, like freezing, fear-induced startle, conditioned flight, etc. In addition, there is remarkable similarity in fear expression and even the neural substrates across individuals and species. Indeed, there is broad consensus on the pivotal role of the amygdala in fear learning in a wide variety of species (see, e.g.,LeDoux,2000) .

Moreover, fear responses are very rapidly acquired, reducing experimental costs tremendously. While the study of many other learning tasks requires lengthy training sessions, significant fear responses can already be observed within few trials. This has contributed to fear conditioning being one of the most well-studied learning paradigms today and one of the earliest fields in which clear links between neural mechanisms and behavior could be established.

Finally, due to the immense importance of the fear system for survival and, hence, high selection pressure, there is good reason to assume it performs in a near-optimal manner. This widens the scope of theoretical approaches tremendously, since it allows for a rational analysis (Anderson,1990) of behavior. That means, considerations pertaining to how information can be optimally processed in the fear circuitry and used to learn to avoid threat are a viable approach to studying fear learning. This will be developed in more detail in chapter 3.

Taken together, in the case of fear learning, it is possible to investigate the nature of the learning process theoretically on at least two levels. On the one hand, a rich literature on the neural substrates is already available and steadily growing, so it is becoming increasingly possible to constrain neurobiological, mechanistic models and derive insight from bottom-up models. On the other hand, it lends itself well to a rational, or normative, analysis, which describes the process from a functional perspective.

1.5 Outline of the Thesis

This thesis is structured as follows: The second chapter is devoted to providing an overview of relevant physiological and anatomical data. This overview reflects the scope of the computational models; it presents the brain structures that have been found to play key roles in the acquisition or extinction of fear responses, outlines their internal microcircuitries and mutual connectivities, and summarizes

(27)

1.5. OUTLINE OF THE THESIS 13

physiological results on the neural activity—and modulation thereof—in the course of fear learning.

The third chapter explains the theoretical background of the high-level modeling approach in more detail. The basic premise of Bayesian learning is introduced and an overview of theoretical models of conditioning in the cognitive sciences is provided. Subsequently, in the fourth chapter, mathematical treatments of neural dynamics are discussed and an approximation for the firing rates of conductance-based integrate-and-fire models is presented and applied to the analysis of dynamics in two-population inhibitory networks.

Chapters five and six constitute the core of this thesis. In them, the two computational models of the fear circuitry are presented and discussed. Finally, the last chapter concludes the work with a discussion of the models, including an analysis of key hypotheses and testable predictions, as well as emerging open questions.

(28)

(29)

Chapter 2 The Neural Substrates of

Fear Learning

This work explicitly aims at providing models that are physiologically constrained. A growing body of experimental literature on fear conditioning and its neural substrates provides the basis for this approach. This research has established that the amygdala, a group of nuclei located in the temporal lobe, is indispensible for the acquisition of conditioned fear responses. For instance, pharmacological lesions of the amygdala lead to a marked decrease in fear aquisition. In addition, the socalled extended amygdala, which includes the central amygdala and stria terminalis, is known to play a key role in mediating anxiety. In particular, the bed nucleus of the stria terminalis is implicated in controlling anxious behavior. Crucially, the neural circuitry involved in the acquisition and extinction of conditioned fear extends much further. The medial prefrontal cortex (mPFC) and hippocampus (HPC) have been reported to shape behavioral expression of both fear and anxiety. Typically, the hippocampus is attributed a pivotal role in contextual modulation of fear responses and the mPFC in high-level control of fear and anxiety. This chapter gives an overview of the neuroanatomy and neurophysiology of fear conditioning and presents results relevant to the theoretical considerations in the main body of this work.

2.1 Basolateral Amygdala

The basolateral complex of the amygdala (BLA) is cosidered the main site of acquisition and storage of fear memories (Davis,1992;Fendt,1999;LeDoux,2000). It can be subdivided into lateral (LA), basal (BA) and accessory basal nuclei. In terms of cytoarchitecture, these nuclei are often described as “cortical”(McDonald,

(30)

16 CHAPTER 2. THE NEURAL SUBSTRATES OF FEAR LEARNING

1992), and accordingly consist of mostly spiny, glutamatergic projection neurons comprising about 80% of the total number of neurons, with an array of different GABAergic interneuron subtypes making up the remainder.

2.1.1 Main Connections

The prominent role of the amygdala in fear conditioning is already apparent in its neuroanatomical structure. Projections from sensory modalities carrying CS-related information and form structures known to transmit nocioceptive signals converge in the BLA, which is the main recipient of external inputs in the amygdala. Specifically, the LA receives sensory inputs from all sensory modal-ities via the cortex and thalamus. These inputs can be subvidived into direct projections from the sensory thalamus (LeDoux,1990) and indirect projections, via the neocortex (LeDoux,1991).

Moreover, the BLA—particularly the BA—is supplied with polymodal inputs from different sources. Most notably, there are inputs from the prefrontal cortex (McDonald,1996; Rosenkranz, 2002), rhinal cortices, and hippocampus (McDonald,1996). A common line of thought is that the prefrontal inputs play a role in mediating behavioral flexibility while the rhinal and hippocampal inputs convey information about context and contextual memory. It is important to note that these connections are reciprocal, indicating a role of the BA in the formation and organization of memory in the mPFC and HPC.

Within the amygdala, connections are directed from the LA to the BA and from both structures to the central amygdala (Ehrlich,2009). Specifically, the LA sends projections to the BA and the capsular division of the CEA. The BA, on the other hand, targets mostly the medial part (CEm) of the CEA. In addition, there are connections to the intercalated cell clusters of the amygdala. The main connections of the BLA are illustrated in Figure 2.1.

2.1.2 Role in Fear Conditioning

A huge body of lesion studies—both permanent and reversible—clearly implicates the BLA as a principal site for the formation and storage of CS-US associations. For instance, it has been shown that lesions of the BLA before conditioning impair acquisition of a fear response, while post-conditioning lesions block expression of the fear response, presumably by preventing the retrieval of the fear memory. Notably, however, some studies using pre-conditioning lesions indicate that the basal part, BA, does not directly contribute to the acqusition and expression of conditioned fear. Fear memory, it was demonstrated, can be acquired and retrieved even in the case of pre- or post-conditioning lesions (Amorapanth,2000;

(31)

2.1. BASOLATERAL AMYGDALA 17

Figure 2.1: Organization of the amygdala circuitry. a) View of an amygdala slice stained with GAD67. The image illustrates the high concentration of GABAergic neurons in the CEA and ITC, as compared to the BLA. b) Simplified scheme of the amygdala circuitry. Sensory inputs reach the LA and are forwarded to the CEA via the BA and ITCs. (adapted fromEhrlich,2009)

CS-dependent Activity and Synaptic Plasticity in the LA

Electrophysiological recording techniques also allow for the investigation of the neuronal activity during fear conditioning. The results corroborated those mentioned before; it was found that the acquisition of a conditioned fear response is accompanied by an increase in CS-evoked activity in the LA. Importantly, this increase is stimulus-specific, i.e., the CS+evokes stronger increases in activity compared to the unpaired CS− (Collins,2000), reflecting the relative rates of conditioned responding.

While such increases could, of course, also be caused by plasticity in afferent structures, e.g., the medial geniculate nucleus of the auditory thalamus (Gerren,

1983), there is ample evidence that they are indeed due to local plasticity within the LA. For example, it could be demonstrated that plasticity in afferent structures is critically dependent on the BLA (Maren,2001). Moreover, there is direct evidence for synaptic plasticity in the LA. Many studies have demonstrated that NMDA-receptor-dependent changes in neuronal activity are essential for the acquisition of conditioned fear responses by local pharmacological interventions (Miserendino,1990;Quirk,1995,1997;Gewirtz,1997;Collins,2000;Rodrigues,

2001). This lends strong support to the notion that NMDA receptor-dependent long-term potentiation in the LA underlies associative learning, establishing a remarkably clear link between synaptic plasticity and observable behavior.

(32)

With respect to plasticity, a line of research that is notable from a theoretical perspective tries to unravel how this synaptic plasticity in the LA is modulated by expectation. Recent results suggest that long-term potentiation in the LA is driven, at least in part, by a sort of reward-prediction error signal that arises in the midbrain periaqueductal gray (PAG) region (McNally,2006,2011). This notion is based on findings that US evoked responses in the LA are stronger for unexpected US than they are for expected US (Belova,2007;Johansen,2010) and that direct stimulation of the PAG can drive fear conditioning (Di Scala,

1987). In line with this, deactivation of the PAG impaired acquisition of a conditioned fear (Johansen,2010). Notably, both the Rescorla-Wagner and the TD learning rules, which will be introduced in section 3.2.1, are based on the concept of expectation modulated learning.

Finally, a number of studies have begun to shed light on the role of inhibitory neurons in the control of synaptic plasticity in the BLA. Activity-dependent potentiation in the LA is facilitated when GABAergic neurons are surpressed (Watanabe,1995;Bissi`ere, 2003; Shaban,2006) and, conversely, activation of GABA-receptors impairs acquisition of conditioned fear (Wilensky,1999). More, recently, it was found that a specific arrangement of two different interneuron subtypes—parvalbumin (PV)- and somatostatin (SOM)-expressing interneurons— plays a crucial role in gating synaptic plasticity in the BLA during fear learning by controlling the activity of the principal neuron bidirectionally (Wolff, 2014). While PV+ neurons preferentially target the soma of the principal neurons and generate feedback inhibition, SOM+ neurons mostly project onto the distal dendrite, and, in addition, the interneurons are differentially recruited by the CS and US. During the CS, PV+neurons are innervated and inhibit SOM+neurons, thereby releasing the principal neuron dendrite from inhibition. Conversely, during the presentation of the US, both interneuron subtypes are inhibited, facilitating principal neuron activity and gating associative plasticity.

Fear and Extinction Neurons in the BA

The discussion so far focused on the acquisition of conditioned fear in the LA. During extinction learning, on the other hand, CS-evoked activity in the LA is decreased (Hobin, 2003; Quirk, 1997) in some neurons—presumably by depotentiation of thalamic inputs (Kim, 2007)—but remains constant in others (Repa, 2001; An, 2012). More remarkably, in the BA, extinction learning is associated with a switch in CS-evoked activity between two subpopulations of principal neurons (Herry,2008;Amano,2011). At the beginning of extinction training, one population displays high CS-evoked phasic activity, correlating with behavioral expression of fear and hence termed fear neurons, but evoked activity

(33)

2.2. INTERCALATED CELL CLUSTERS 19

gradually decreases in the course of extinction learning. Curiously, another population, called extinction neurons, behaves in the exact opposite way: there is little to no evoked activity at the beginning, but the neurons acquire CS-evoked responses during extinction. This switch in neural activity precedes the decline in conditioned responding (see figure 2.2). Finally, a third population of principal neurons is resistant to extinction learning, i.e., they exhibit CS-evoked phasic activity throughout extinction learning. Notably, this switching between fear and extinction neurons echoes the idea of fear and extinction memory traces that was proposed based on behavioral results, most prominently the phenomenon of fear renewal.

Mechanistically, the activity of fear and extinction neurons indicates mutual competition. This led to the hypothesis that the switching is mediated by intra-BA inhibitory neurons. In line with this, an increase in GABA levels after extinction learning (Heldt,2007) can be observed, and there is an increase in IPSC amplitude and frequency in BA principal cells after extinction (Lin,

2009). Adding to this, a recent study reported differential plasticity of inhibitory synapses depending on whether the cells targeted fear neurons, displaying a decrease in evoked activity during extinction, or extinction-resistant neurons (Trouche,2013).

Importantly, interfering with this microcircuitry blocks behavioral transitions, but not specifically expression of conditioned fear or fear extinction(Herry,2008). Injecting the GABA-agonist muscimol into the BA at different time points in the paradigm has the effect of blocking transitions between high-fear and low-fear states, e.g., blocking fear extinction during safety learning or fear renewal when changing context (see Figure 2.2). This implies a role of the BA in modulation and control of fear, while the LA appears as the main locus of associative learning.

2.2 Intercalated Cell Clusters

Further, the intercalated cells (ITC) of the amygdala have been implicated in fear extinction. The ITCs do not form a cohesive nucleus, but rather a number of small, densely packed clusters of mostly GABAergic (Par´e,1993) cells around the BLA (see figure 2.3). Based on their position relative to the BLA, they are usually divided into lateral ITCs (lITC), medial ITCs (mITC), and the baso-medially located main cluster (ITC) (Ehrlich,2009). They are well connected within the amygdala (Geracitano,2007;Millhouse,1986;Royer,1999), with the lITC exerting inhibitory control of the BLA (Marowsky,2005), while the mITCs and ITCs gate information flow from BLA to CEA (Par´e,2003;Royer,1999). The ITCs receive sensory input from the thalamus and cortex (Asede,2015), and dense connections from the infralimbic cortex (Millhouse,1986;McDonald,

(34)

Figure 2.2: Fear and extinction neurons. a) During extinction learning, fear neuron responses gradually decrease, while CS-evoked activity in extinction neurons increases. Freezing responses diminish after the switch in neural activity (gray bars). b) Behavioral tran-sitions are blocked by selective and re-versible inactivation of the BA. Top row: inaction after training prevents acquisition of extinction. Bottom row: inactivation after extinction pre-vents fear renewal. (adapted from Herry,

2008)

1996;Vertes,2004), which can cause strong excitation of the ITCs (Amir,2011). Their inter-amygdala connections are organized topographically; the mITCs receive projections mostly from principal cells in the LA, while the ITCs are targeted by BA principal neurons, and synapse onto adjacent CEA neurons (Par´e,2003;Royer,1999). Moreover, there is substantial intra-cluster recurrent

connectivity (Geracitano,2007,2012).

This connectivity already points towards a role in controlling CEA excitabil-ity and hence fear expression, and indeed ITCs are mostly implicated in fear extinction learning (Par´e,2003). Extinction training leads to increased activity in the ITC, as evidenced by heightened c-fos and Zif628 expression (Knapska,

2009;Busti,2011). Moreover, it can be demonstrated that mITCs are necessary for the expression of fear extinction memory by selective lesion (Likhtik,2008), or conversely, that facilitation of ITC activity enhances fear extinction (J¨ungling,

2008). Lastly, extinction training is accompanied by potentiation of BA synapses onto ITC, presumably inhibiting CEm (Amano, 2010). Notably, this effect is dependent on activity in the infralimbic cortex (ibid.).

(35)

2.3. CENTRAL AMYGDALA 21

fear expression. It was demonstrated that BLA-mITC connections undergo potentiation already during fear learning and that inputs from sensory areas exhibit plasticity as well (Asede,2015). These results indicate that ITCs also induce fear expression via disinhibition of the CEm (Busti,2011) and suggests that, instead of just inhibiting fear expression, ITCs might form a parallel pathway to LA that is capable of noth promoting and inhibiting fear expression.

2.3 Central Amygdala

The central amygdala (CEA) is a GABAergic nucleus located dorsomedially with respect to the basolateral complex. Anatomically and physiologically, the CEA can be subdivided into a lateral (CEl) and a medial (CEm) nucleus. Functionally, it is generally considered the main output region of the amygdala and plays a pivotal role in fear expression. While it was long regarded as a mere passive relay in the fear circuitry, recent research highlights its role in acquisition of fear responses and particularly fear generalization.

2.3.1 Connections with Other Brain Structures

In the fear pathway, the CEA is the next structure downstream of the basolateral complex receiving amygdala-internal projections from the BLA (Pitk¨anen,1995), as well as the ITCs. Moreover, it receives direct projections from sensory areas (Sah,2003) including the auditory thalamus (Samson,2005). Complementing these, the CEA receives nocioceptive input as well via connections from the parabrachial nucleus and solitary tract (Shimada,1992;Jhamandas,1996;Dong,

2010).

Moreover, the CEA has abundant out-bound projections to other brain regions. The medial part consists of neurons targeting the hypothalamus (LeDoux

,1988) and various brainstem nuclei (Veening,1984). Of particular relevance for the freezing response typically observed in the conditioning paradigm are the connections to the periaqueductal gray (Behbehani, 1995; Rizvi, 1991), a structure known to mediate analgesia (Basbaum, 1984) and defensive responses like freezing (LeDoux, 1988; Davis, 1992). These different output pathways mediate distinct behavioral fear responses (LeDoux,1988;LeDoux,2000).

2.3.2 Internal Structure: CElon and CEloff

As for internal structure, there are intrinsic connections (Jolkkonen,1998;Lopez de Armentia, 2004), and the wealth of neuron subtypes in the CEA (Viviani,

2011; Veinante,1997) points towards the importance of inter-CEA inhibition (Veinante, 2003; Huber, 2005; Ehrlich, 2009). Recent studies (Ciocchi, 2010;

(36)

Haubensak,2010) revealed and characterized a specific functional microcircuitry within the CEl of particular importance to conditioning. During conditioning, two subpopulations become distinguishable by their responses to the CS: one exhibits excitatory responses (termed CElon), while the other population gets inhibited (termend CEloff). The medial nucleus CEm, in turn, increases its activity on CS presentation. Overall, the picture of an inhibitory microcircuitry emerges, where the CElon subpopulation gets innervated by CS input from the BLA and thalamus and inhibits the CEloff population by direct synaptic connections. As a consequence, the CEm is released from inhibition, leading to freezing (Figure 2.3). Notably, this functional distinction in CElon and CEloff coincides with the expression of the protein kinase PKCδ (Haubensak, 2010). CEloff neurons, i.e., the subpopulation of CEl neurons inhibited by the CS after conditioning, expresses PKCδ, while CElon neurons do not. This microcircuitry is illustrated in Figure 2.3 a.

2.3.3 Synaptic Plasticity in the CEA

Already before the discovery of this microcircuitry, studies have increasingly pointed towards active changes in the CEA during fear conditioning. For instance, reversible pharmacological interference in the CEA during fear conditioning was reported to reduce fear responses during testing (Wilensky,2000;Goosens,2003) and it was found that fear responses can be acquired by overtraining after BLA lesions, a process that is CEA-dependent (Zimmerman, 2007; Rabinak,

2008). More recent results (Li,2013;Watabe,2013;Penzo,2014) provide direct evidence for synaptic potentiation and depression and, importantly, indicate that plasticity within the CEl is subpopulation-specifc. The connections from the BLA to SOM+ CEl Neurons, which roughly overlap with CElon neurons, show a tendency to increase synaptic efficacy, while connections to SOM- neurons, overlapping with CEloff, tend to decrease. This switch in relative synaptic efficacy facilitates acquisition of a CS-evoked network response.

2.3.4 Tonic Inhibition in the CEA

Another important aspect of neural plasticity in the CEA relates to the tonic activity. A salient finding in Ciocchi (2010) was that not only did phasic, CS-evoked activity in the CEA change during conditioning, but also tonic activity, i.e., the baseline firing, changed with experience. In CElon and CEm neurons, baseline firing rate tends to decrease, while in CEloff, it increases. Remarkably, the magnitude of these changes in tonic activity relates to the behavioral expression of fear generalization. Animals that displayed stronger increases in CEloff rate tended to generalize, i.e., exhibit higher CS− firing.

(37)

2.3. CENTRAL AMYGDALA 23

Figure 2.3: The CEA microcircuitry. a) Sketch of the disinhibitory CEA microcircuitry. b) Top row: Phasic responses for the three CEA subpopulations before and after conditioning. Bottom row: Correlation with fear generalization. Particularly in CEloff, there is a strong positive correlation between tonic rate increase and fear generalization score. (adapted fromCiocchi, 2010)

Following up on these results, PaoloBotta (2015) showed that CEA neurons undergo modulation of tonic inhibition during fear learning. Tonic inhibition denotes persistent currents mediated by extrasynaptic GABAAreceptors and

has been reported in other brain areas previously (Kaneda,1995;Nusser, 2002;

Semyanov, 2004). These have a different structural composition and different properties from their synaptic counterparts, most importantly a higher affinity for GABA and low receptor desensitization (Farrant,2005). By virtue of these properties, they are persistently activated by low concentrations of GABA and mediate a tonic inhibitory current on the cell membrane.

Importantly, in PKCδ+_{neurons in the CEA, these tonic currents decrease.}

This is fully consistent with the increase in baseline firing of CEloff neurons reported previously. Furthermore, the effects on fear generalization are also consistent: the lower the tonic inhibition in PKCδ+ neurons, the higher the fear generalization scores. Critically, this is not a mere correlation; optogenetic manipulation of the PKCδ+ _{population modulates fear generalization in the}

same way. This lends strong support to the idea that tonic inhibition in the CEA controls fear generalization.

Relation to Anxiety

Just like fear, anxiety is mediated by a distributed circuitry in which both the BLA and the CEA are involved (Tovote, 2015). Early studies implicated the CEA in the control of anxiety (Jellestad, 1986) and, more recently, it has been shown that GABAergic signalling in the amygdala affects anxiety (Tasan,

(38)

Figure 2.4: Tonic inhibition in the CEA. a) Extrasynaptic inhibition in PKCδ+ decreases during fear learning. The right panel shows example current traces. b) Fear generalization correlates with the post-FC tonic inhibition (left panel) and stimulation of PKCδ+ cells increases fear generalization. c) Optogenetic manipulation can modulate anxiety in the elevated plus maze (left panel) and open field test (right panel) bidirectionally. (adapted from Botta,

2015)

2011). Together with the relation between fear generalization and anxiety, this points towards a role of tonic inhibition in the central amygdala in the control of anxiety. Indeed, it could be demonstrated that optogenetic stimulation of PKCδ+ neurons increases anxiety scores in the open field test and elevated plus test, while inhibition reduces them (Botta,2015).